By George Burwood, Ph.D., and Julia Dietlmeier, Ph.D.
Cochlear implants (CIs) are advanced hearing devices that restore hearing by sending electrical signals to the auditory nerve. Some implants, called hybrid CIs, combine electrical stimulation with the patient’s remaining residual low frequency hearing.
However, scar tissue (fibrosis) can form inside the inner ear after implantation, which may interfere with the residual hearing and reduce the benefits of hybrid CIs over time. There is ample evidence that the scarring is a result of the inner ear’s immune response to implantation.
While many strategies are being developed to solve this problem, the way inner ear scarring affects acoustic hearing is not fully described. Identifying how much, where, and what type of scarring exists inside implanted cochleae, and comparing those features to residual hearing function, has implications for device design, surgical approach, patient counseling, stimulation paradigms, and implantation criteria.
To improve outcomes for CI users, we are studying how fibrosis forms in an animal model by embracing the power of artificial intelligence to aid in imaging analysis. As detailed in our paper in the IEEE journal Transactions in Biomedical Engineering, published in January 2025, we applied a computer vision approach to process optical coherence tomography (OCT) images of chronically implanted rodent cochleae, in order to better understand how and when residual hearing loss occurs after implantation.
This is the first project to measure vibrations in chronically implanted cochleae using OCT, and to produce the first OCT imaging dataset of implanted cochleae. OCT can be thought of as ultrasound imaging but using infrared light. Routinely used to image the eye, it has properties that make it an important tool for inner ear mechanics research.
A New Computer Vision Model
Computer vision aided by machine learning is a powerful tool when scientists want to identify structures in a complicated medical scan, in large numbers, and with consistency. In this work we designed our computer vision model, 2D-OCT-UNET, to process the OCT imaging data from the experiments.
This model is a development of the UNET type of neural network. Named after its U-shaped architecture, UNET allows sharing of image features at different scales, from patterns in neighboring pixels to larger shapes.
A well-known feature of UNET is that it can work well with little training data. This feature is vital because the number of images available to train the network is limited, although is increasing thanks to the tireless efforts of a group of Oregon Health & Science University medical students (among the study’s co-authors) who volunteered to manually annotate a subset of the image database.
We added several modifications to the basic UNET architecture. First, we increased the depth of the network, which facilitates extraction and processing of more semantically coherent visual features and permitted the model to assess higher resolution images.
Then we added more processes within the algorithm designed to extend its robust response to a smaller training dataset. It should be noted that even a “small” training dataset still contains hundreds of images and many thousands in the overall dataset. This is because each OCT scan builds a 3D image of the inner ear, the fibrosis, and the implant, at microns-resolution scale.
Because of the increased network depth, upscaled input resolution, and the ability to perform well in a limited data setting, our 2D-OCT-UNET outperformed state-of-the-art models on our cochlear OCT dataset in tasks such as identifying inner ear scarring, CI electrode track position, and the fluid spaces of the inner ear.
We are continuing with this work by increasing the size of the training dataset and setting the UNET other tasks, such as comparing which structures within the inner ear are afflicted by scarring.
We hope that our findings will advance future studies on exploring the relationship between cochlear fibrosis and residual hearing loss, the development of cochlear implants, and the treatment of patients receiving electric-acoustic stimulation, a treatment for patients who are profoundly deaf in the high frequency region but retain usable low frequency hearing.
George Burwood, Ph.D., is a research instructor at Oregon Hearing Research Center, Oregon Health & Science University. He is a 2023 Emerging Research Grants (ERG) scientist. The study’s first author Julia Dietlmeier, Ph.D., a computer vision expert, is a senior postdoctoral researcher at the Insight Research Ireland Centre for Data Analytics at Dublin City University in Ireland. Coauthor Lina Reiss, Ph.D., is a 2012–13 ERG scientist.
Because of the increased network depth, upscaled input resolution, and the ability to perform well in a limited data setting, our new imaging model aided by machine learning outperformed previous state-of-the-art models predicting cochlear implant success outcomes.